Aligning Human and Machine Attention for Enhanced Supervised Learning

Ahmed Abbasi; Avihay Chriqui; Dov Teeni; Inbal Yahav

REVIEW 4 major objections 6 minor 15 references

Aligning Human and Machine Attention for Enhanced Supervised Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read HuMAL-AL, a regularizer that aligns BERT's last-layer attention with human word highlights, improves text classification AUC when labels are sparse and imbalanced.

desk verdict Equation 2 as printed optimizes against human attention, not toward it, so the paper's central claim is unsupported as written — though the empirical setup is extensive and the fix may be straightforward. read the letter →

arxiv 2502.06811 v2 pith:6A4BWKBU submitted 2025-02-04 cs.LG cs.AIcs.CL

Avihay Chriqui , Inbal Yahav , Dov Teeni , Ahmed Abbasi This is my paper · ORCID

classification cs.LGcs.AIcs.CL

keywords humanattentionmachinealignmentBERTfine-tuningimbalancedtextclassificationsentimentanalysispersonalitysupervisedlearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that adding a human-attention alignment term to the fine-tuning loss of a transformer improves classification, and that the gain is largest exactly where data is hardest: 1–5% class imbalance and only 250–500 labeled instances. On Yelp sentiment, the HuMAL-AL variant raises mean AUC from 0.64 to 0.84 at 250 instances and 1% imbalance, and from 0.75 to 0.94 at 500 instances; it also beats fine-tuned BERT on introvert–extrovert classification of myPersonality posts. The paper reads this as evidence that cheap, task-specific human highlights can stand in for larger labeled datasets in low-resource text classification. A sympathetic reader would care because the annotations are simple word-level highlights from three crowd workers, not eye-tracking or brain data, and the same recipe transfers to GPT-2 and XLNet.

What carries the argument

The load-bearing object is the HuMAL-AL loss, written in the paper as $L = \sum_j \mathrm{cross\_entropy}(y_j, p_j) + \alpha \cdot \cos(A^h_j, A^m_{j,11})$, with $\alpha=2$. The machine attention vector $A^m_{j,11}$ comes from averaging, over the 12 heads, the CLS token's attention to each token in BERT's last layer (Equation 1); the human vector $A^h_j$ is the sum of binary highlight decisions across annotators, normalized to sum to 1 (Equation 5). This couples the classification objective to where the model looks, rather than only what it predicts. The paper also defines two alternatives: HuMAL-AN blends human attention into the final embedding, and HuMAL-AP injects human attention as a prior on the first layer instead of the last.

What would settle it

Train BERT on a small Yelp subset with the loss exactly as printed in Equation 2 and log the cosine similarity between human and machine attention at each epoch: if the equation is literal, similarity should decrease toward -1, contradicting the paper's reported rise from 0.27 to 0.45. Inspecting the released training code for the sign of the attention term would settle it directly.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that HuMAL-AL—the strategy that adds a cosine-similarity term between the human attention vector and BERT's last-layer attention to the classification cross-entropy—improves AUC over fine-tuned BERT, and that the gain is largest at 1–5% imbalance with 250–500 training instances. The authors report this advantage on Yelp sentiment and myPersonality introvert/extrovert classification, across BERT, GPT-2, and XLNet backbones. They also report that HuMAL-AL reaches baseline-level AUC with roughly 484 fewer labeled reviews at 5% imbalance, and that cosine similarity between human and machine attention rises from about 0.27 to 0.45. The explanation offered is that task-specific human attention marks sentiment- and personality-relevant words, and the model is regularized toward those marks.

Load-bearing premise

The reported gains depend on the deployed attention loss actually maximizing similarity to human attention, whereas Equation 2 as printed adds the cosine term and would push similarity to -1.

Editorial extensions

If this is right

At 1% imbalance and 250 training instances on Yelp sentiment, HuMAL-AL raises mean AUC from 0.64 (BERT) to 0.84; at 500 instances it raises it from 0.75 to 0.94.
At 5% imbalance, HuMAL-AL matches BERT while needing about 484 fewer labeled reviews on average.
The gains transfer to GPT-2 and XLNet as base models, with the largest improvements again at 1% imbalance and small sample sizes.
On personality classification from myPersonality posts, HuMAL-AL improves balanced AUC from 0.68 to 0.74 with majority-vote labels and from 0.64 to 0.75 with self-reported labels.
The benefit shrinks with text length: for 101–200 word reviews, the gains at small sample sizes are smaller than for reviews up to 50 words.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the sign of the attention term is corrected, the mechanism is a soft constraint that pulls the model's focus onto the same roughly 13–21% of words humans mark, suggesting that a three-highlighter annotation protocol could substitute for thousands of labels in other low-resource text tasks.
The paper fixes the alignment target to the last layer, and its own first-layer variant HuMAL-AP performs worse, so a natural extension is to test whether the last layer is the best alignment point or whether layer choice should depend on task depth.
Because human attention agreement drops on longer texts and HuMAL's gain drops too, the regularization weight $\alpha$ could be made to decay with text length or annotator agreement, a design the paper does not test.
The personality results are noisy (AUC standard deviations around 0.1), so a follow-up with more instances and a pre-registered sign for the attention loss would clarify whether the reported effect is robust.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my read. The paper has a promising premise and a large, careful evaluation, but the central math as printed contradicts the reported results. Equations 2 and 4 add alpha times cosine similarity to the cross-entropy. Minimizing that sum drives cosine similarity toward -1, i.e., away from human attention. Yet Table V reports cosine similarity increasing from 0.27/0.19 to 0.45-0.48 under HuMAL-AL. That is not possible under the stated objective. No code is released to show whether the experiments used a corrected loss, so the main mechanism is unsupported as written.

What the paper does well: it systematically compares three distinct integration strategies (loss, normalizer, prior) for task-specific self-reported attention, across three architectures and two tasks. The low-resource and imbalanced framing is practical, and the extra analyses on text length and loss convergence are thoughtful. If the loss sign is a typo, this could become a genuinely useful empirical contribution.

The soft spots go beyond the sign error. Eq. 3 mixes attention vectors with the embedding vector in a way that isn't dimensionally consistent, and the 'simplification' drops alpha and the machine-attention term. The authors cite McGuire and Tomuro (2021) and Zou et al. (2018) but do not benchmark against them. Table V's similarity measure uses the same annotations as the training signal, so part of that increase is expected, not independent validation. And in the personality task, HuMAL-AL only improves over BERT in the balanced setting, which undercuts the headline about imbalance.

The central flaw is load-bearing. It is not a minor equation typo you can wave away, because the reported alignment gain would be impossible under the printed loss. That said, the rest of the paper looks like honest empirical work, and the fix is plausibly simple (change the sign or use 1-cos), with code and the missing baselines added.

I would send this to peer review rather than desk reject — a referee can force the equations to match the implementation and require artifacts. For my own citing purposes, not in current form. Reading group? Maybe, if you want a concrete example of a results section and an optimization objective contradicting each other.

Referee Report

4 major / 6 minor

Summary. The paper introduces HuMAL, three strategies for injecting human word-level attention annotations into transformer fine-tuning, and evaluates them on Yelp sentiment analysis and myPersonality introvert/extrovert classification. The flagship variant HuMAL-AL adds a cosine-similarity term between human attention and BERT's last-layer attention to the cross-entropy loss; the paper reports that HuMAL-AL improves AUC over fine-tuned BERT, especially with 1–5% imbalance and 250–500 training instances, and increases human–machine attention cosine similarity from about 0.19–0.27 to 0.45–0.48. The other two variants, HuMAL-AN and HuMAL-AP, give smaller or no gains. The paper also presents data-efficiency analyses and GPT-2/XLNet replications.

Significance. If the results are valid, the work would offer a cost-effective way to leverage cheap self-reported highlights to improve low-resource classification, and it would contrast with earlier negative results from eye-tracking/EEG and lexicon-based guidance. Strengths include the use of existing human-attention data on Yelp (Sen et al., 2020), comparison across three transformer architectures, replication across two tasks, and bootstrapped significance testing. However, the central objective as printed cannot produce the reported attention-alignment results, so the significance is conditional on a substantial correction and on code release.

major comments (4)

[Section III-A, Eq. (2) and Section III-C, Eq. (4)] The loss is written as L = Σ_j CE(y_j, p_j) + α·cos(A^h_j, A^m_{j,11}), with α=2 stated in Section IV-C. Since training minimizes the loss, the plus sign makes the cosine-similarity term push cos toward −1, i.e., it actively penalizes alignment with human attention. This contradicts the design statement in Section III that the model is 'penalized for deviating from human attention,' and it makes the increases in Table V (from 0.27/0.19 to 0.45–0.48) impossible under the printed loss. If the implemented objective was instead CE − α·cos or CE + α·(1−cos), the equations misstate the method, and no code or supplementary material is provided to resolve the ambiguity. This is a load-bearing internal inconsistency for the central mechanism.
[Section III-B, Eq. (3)] The first expression states E_{j,11}^{AN} = (A^m_{j,11} + α A^h_j) × E_{j,10}, but the second equality gives E_{j,11} + A^h_j × E_{j,10}, dropping both the α coefficient and the role of A^m_{j,11} in the decomposition. If A^m_{j,11} E_{j,10} is meant to be E_{j,11}, the α should still appear in the second term. The equation also does not specify whether A^h_j is normalized to match the scale of A^m_{j,11}. As written, Eq. (3) is algebraically incorrect.
[Section V-A, Table III(b)] The text claims the performance benefit is 'particularly pronounced' under high class imbalance and small datasets, but in the personality task the imbalance rows show no significant improvement for HuMAL-AL over BERT (0.59 vs 0.57 and 0.62 vs 0.60, neither starred). The claim is only supported by the sentiment task and by the balanced personality condition; the summary and abstract should be scoped accordingly.
[Section V-E, Table V] The attention similarity reported for HuMAL-AL is computed against the very same human annotations that were used as the training target in the HuMAL-AL loss. The increase from BERT to HuMAL-AL is therefore a direct consequence of the objective and does not constitute independent evidence of human-like attention. It should be described as a manipulation check, with the held-out label AUC in Tables III and VI serving as the primary evidence.

minor comments (6)

[Table VI] The table uses 'HuMAL3-xlnet' in the header and rows, while the text and other tables use 'HuMAL-AL-xlnet'; please standardize the naming.
[Table IV] The table has layout problems: Yelp-100 and Yelp-200 rows contain misplaced 'Not enough data' cells and missing or duplicated entries, making the results difficult to parse.
[Eq. (5)] The normalization formula writes 'Pi i=0' as a product symbol; it should be a summation over i, i.e., Σ_i.
[Section V-A] The text states results were consistent 'regardless of whether human attention was considered binary (first labeler only) or continuous,' but no comparison of binary vs. continuous conditions is reported in any table; either add the analysis or remove the claim.
[Section IV-B] The test split and test-set size for the myPersonality task are not specified; Table II only lists training instances. Please clarify how the evaluation set was constructed and how the imbalance/balance categories map to it.
[Throughout] There are several typographical errors, including 'Our goal in aligning machine attention ... to enhance' missing a verb in Section II and an extra period after 'performance..' in Section VI.

Circularity Check

1 steps flagged · score 3.0 of 10

Attention-similarity evaluation is self-referential, but the central AUC claim is tested on held-out labels and is not circular.

self definitional [Section III-A, Eq. (2); Section V-E, Table V]
"AL loss = Σ_j cross entropy(y_j, p_j) + α × cosine similarity(A^h_j, A^m_{j,11}). Table V: BERT 0.27 (0.10) 0.19 (0.02) ... HuMAL-AL 0.45* (0.02) 0.47* (0.02)."

The evidence that HuMAL-AL 'shifted the machine's attention closer to that of human annotators' is the same cosine similarity that the HuMAL-AL objective is designed to optimize. Table V reports an increase in that exact quantity from 0.27/0.19 to 0.45-0.48, so the improvement is the training target itself, not an independent confirmation that aligning to human attention improves classification. The paper does not state that the attention-similarity analysis uses human annotations held out from training, and Section IV-C says the human attention vectors 'were used only for training.' If the values in Table V come from the training annotations, the rise is expected by construction. Independently, the printed plus sign in Eq.

full rationale

The central performance claims are not circular: the AUC improvements in Tables III, IV, and VI are computed on held-out test sets (e.g., a fixed balanced test set of 200 instances) and compare HuMAL against fine-tuned BERT, GPT-2, and XLNet, so they stand as external benchmark results. The one self-referential element is Table V's attention-similarity result, which measures the exact cosine similarity term appearing in the HuMAL-AL training loss (Eq. 2). Since human attention vectors were used as the training signal, a rise in that metric on those same annotations is a property of the objective rather than evidence that the mechanism is validated. The self-citations in the paper (Te'eni et al. 2023; Zagalsky et al. 2021) are motivational and do not carry any load-bearing technical inference. There is also an internal inconsistency in Eq. (2): minimizing the printed loss would decrease cosine similarity, contradicting Table V; that is a correctness problem, not a circularity, but it further weakens the attention-alignment narrative. On balance, the main AUC derivation is self-contained and externally evaluated, so the circularity score is moderate rather than high.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central method uses no new physical or architectural entities; HuMAL is a training procedure. The main uncharged assumptions are the validity of averaged attention as an importance measure and the correctness of the loss sign.

free parameters (2)

alpha (regularization coefficient) = 2
Chosen by hand and fixed to 2 for all HuMAL variants; the balance between classification and attention terms depends on it, and no sensitivity analysis is given.
minimum annotation coverage threshold = 2%
Instances with fewer than 2% of words annotated were excluded in the myPersonality task (Section IV-B); this affects which instances appear in training.

assumptions (4)

domain assumption Attention scores averaged over all heads in a BERT layer provide a token-level importance vector comparable to human word highlights.
Equation 1 and Section III average over 12 heads following Clark et al. [2019]; no validation is given that this average is a faithful importance measure.
domain assumption Cosine similarity between normalized human binary attention and continuous machine attention is a valid alignment objective.
Equations 2 and 4 use cosine similarity as the attention loss; the paper does not compare alternative alignment metrics.
domain assumption Three annotators per instance yield homogeneous task-specific human attention without need for inter-annotator agreement statistics.
Section IV-B uses at least three annotators; the Discussion claims 'remarkable homogeneity' but no agreement numbers are reported.
ad hoc to paper The implemented attention loss maximizes similarity to human attention despite the printed plus sign in Eq. 2.
Needed to reconcile Eq. 2 with the stated goal and the increased similarity in Table V; without code this is an unverified assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aligning Human and Machine Attention for Enhanced Supervised Learning." pith.science (2026). https://pith.science/paper/6A4BWKBU

@misc{pith2026250206811,
  author       = {Pith},
  title        = {Pith review of: Aligning Human and Machine Attention for Enhanced Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6A4BWKBU}},
  note         = {Machine review of arXiv:2502.06811}
}

read the original abstract

Attention, or prioritization of certain information items over others, is a critical element of any learning process, for both humans and machines. Given that humans continue to outperform machines in certain learning tasks, it seems plausible that machine performance could be enriched by aligning machine attention with human attention mechanisms -- yet research on this topic is sparse and has achieved only limited success. This paper proposes a new approach to address this gap, called Human-Machine Attention Learning (HuMAL). This approach involves reliance on data annotated by humans to reflect their self-perceived attention during specific tasks. We evaluate several alternative strategies for integrating such human attention data into machine learning (ML) algorithms, using a sentiment analysis task (review data from Yelp) and a personality-type classification task (data from myPersonality). The best-performing HuMAL strategy significantly enhances the task performance of fine-tuned transformer models (BERT, as well as GPT-2 and XLNET), and the benefit is particularly pronounced under challenging conditions of imbalanced or sparse labeled data. This research contributes to a deeper understanding of strategies for integrating human attention into ML models and highlights the potential of leveraging human cognition to augment ML in real-world applications.

Figures

Figures reproduced from arXiv: 2502.06811 by the authors.

**Figure 2.** HuMAL illustration: Alternative strategies for lever [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison between the performance of HuMAL-AL [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Performance of HuMAL-AL compared to BERT across [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: HuMAL-AL: Loss Convergence Analysis For the personality type classification task, we split the myPersonality dataset into two groups based on text length: short texts (up to 100 words) and longer texts (101 words and above). Similarly to what we observed in the Yelp an…

**Figure 6.** Figure 6: Average attention as a function of token position [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 5 canonical work pages

[1]

Deriv- ing machine attention from human rationales

Yujia Bao, Shiyu Chang, Mo Yu, and Regina Barzilay. Deriv- ing machine attention from human rationales. arXiv preprint arXiv:1808.09367,

arXiv
[5]

Using convolutional neural network with bert for intent determination

Changai He, Sibao Chen, Shilei Huang, Jian Zhang, and Xiao Song. Using convolutional neural network with bert for intent determination. In 2019 International Conference on Asian Language Processing (IALP) , pages 65–70. IEEE,

work page 2019
[6]

Coca: Cost-effective collaborative annotation system by combining experts and amateurs

Jiayu Lei, Zheng Zhang, Lan Zhang, and Xiang-Yang Li. Coca: Cost-effective collaborative annotation system by combining experts and amateurs. In 2022 IEEE 38th In- ternational Conference on Data Engineering (ICDE) , pages 674–685. IEEE,

work page 2022
[9]

A comprehensive overview of large language models

Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 ,

arXiv
[10]

Llama: Open and efficient foundation language models

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 ,

arXiv
[11]

From Characters to Words to in Between: Do We Capture Morphology?

Clara Vania and Adam Lopez. From characters to words to in between: Do we capture morphology? arXiv preprint arXiv:1704.08352,

work page Pith review arXiv
[12]

Analyzing the structure of attention in a transformer language model

Jesse Vig and Yonatan Belinkov. Analyzing the structure of attention in a transformer language model. arXiv preprint arXiv:1906.04284,

arXiv 1906
[13]

Huggingface’s transformers: State-of-the-art natural language processing

Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R’emi Louf, Morgan Funtowicz, et al. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771 ,

arXiv 1910

Show all 15 references

[14]

Human attention during goal-directed reading comprehen- sion relies on task optimization

Jiajie Zou, Yuran Zhang, Jialu Li, Xing Tian, and Nai Ding. Human attention during goal-directed reading comprehen- sion relies on task optimization. bioRxiv, pages 2023–04,

2023
[15]

A lexicon-based supervised attention model for neural senti- ment analysis

Yicheng Zou, Tao Gui, Qi Zhang, and Xuan-Jing Huang. A lexicon-based supervised attention model for neural senti- ment analysis. In Proceedings of the 27th international conference on computational linguistics , pages 868–877, 2018

2018
[2017]

Effective approaches to attention-based neural machine translation

Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025 ,

arXiv
[2019]

Evo- lutionary data measures: Understanding the difficulty of text classification tasks

Edward Collins, Nikolai Rozanov, and Bingbing Zhang. Evo- lutionary data measures: Understanding the difficulty of text classification tasks. arXiv preprint arXiv:1811.01910,

arXiv
[2021]

Bert: Pre-training of deep bidirectional trans- formers for language understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. arXiv preprint arXiv:1810.04805,

arXiv
[2022]

Rationalizing neural predictions

Tao Lei, Regina Barzilay, and Tommi Jaakkola. Rationalizing neural predictions. arXiv preprint arXiv:1606.04155,

arXiv
[2023]

What does bert look at? an analysis of bert’s attention

Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christo- pher D Manning. What does bert look at? an analysis of bert’s attention. arXiv preprint arXiv:1906.04341 ,

1906 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Deriv- ing machine attention from human rationales

Yujia Bao, Shiyu Chang, Mo Yu, and Regina Barzilay. Deriv- ing machine attention from human rationales. arXiv preprint arXiv:1808.09367,

arXiv

[2] [5]

Using convolutional neural network with bert for intent determination

Changai He, Sibao Chen, Shilei Huang, Jian Zhang, and Xiao Song. Using convolutional neural network with bert for intent determination. In 2019 International Conference on Asian Language Processing (IALP) , pages 65–70. IEEE,

work page 2019

[3] [6]

Coca: Cost-effective collaborative annotation system by combining experts and amateurs

Jiayu Lei, Zheng Zhang, Lan Zhang, and Xiang-Yang Li. Coca: Cost-effective collaborative annotation system by combining experts and amateurs. In 2022 IEEE 38th In- ternational Conference on Data Engineering (ICDE) , pages 674–685. IEEE,

work page 2022

[4] [9]

A comprehensive overview of large language models

Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435 ,

arXiv

[5] [10]

Llama: Open and efficient foundation language models

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 ,

arXiv

[6] [11]

From Characters to Words to in Between: Do We Capture Morphology?

Clara Vania and Adam Lopez. From characters to words to in between: Do we capture morphology? arXiv preprint arXiv:1704.08352,

work page Pith review arXiv

[7] [12]

Analyzing the structure of attention in a transformer language model

Jesse Vig and Yonatan Belinkov. Analyzing the structure of attention in a transformer language model. arXiv preprint arXiv:1906.04284,

arXiv 1906

[8] [13]

Huggingface’s transformers: State-of-the-art natural language processing

Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R’emi Louf, Morgan Funtowicz, et al. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771 ,

arXiv 1910